Description
Classifying the Geolocation of Tweets
In this assignment, you will develop and critically analyse geolocation classifiers for Tweets. That is, given a tweet, your model(s) will produce a prediction of where the author of the tweet was located. You will be provided with a data set of tweets that have been annotated with the location of the author. The assessment provides you with an opportunity to reflect on concepts in machine learning in the context of an open-ended research problem, and to strengthen your skills in data analysis and problem solving.
The goal of this assignment is to critically assess the effectiveness of various Machine Learning classification algorithms on the problem of determining a tweeter’s location, and to express the knowledge that you have gained in a technical report. The technical side of this project will involve applying appropriate machine learning algorithms to the data to solve the task. There will be a Kaggle in-class competition where you can compare the performance of your algorithms against your classmates.
The focus of the project will be the report, formatted as a short research paper. In the report, you will demonstrate the knowledge that you have gained, in a manner that is accessible to a reasonably informed reader.
Data Sets
You will be provided with a geolocation-labelled training set of Tweets, a geolocation-labelled development set which you can use for model selection and tuning, and an unlabeled test set which will be used for final evaluation in the Kaggle in-class competition. Train, test and development sets are provided for each of the representations explained below. Each row in the data files contains the twitter user ID and the tweet representation as comma-separated values. Each line in the label files contains the corresponding location label.
Class Labels
In the provided data set, each tweet is labelled with one of four possible regions (i.e., class labels):
[MIDWEST, NORTHEAST, SOUTH, WEST]
Features
To aid in your initial experiments, we have applied some feature engineering to the raw tweets. You may use any subset of the representations described below in your experiments, and you may also extract your own features from the tweets if you wish. The provided representations are
- full The raw tweets represented as a single string, e.g.,
User ID,“ill explain everything in a bit ill explain you everything”
- count which stands for “Bag of Words”. We (1) filtered the words in the data set, removing very frequent and very infrequent words; and (2) mapped each word to a unique ID. The resulting mapping from remaining word strings to their ID is provided in vocab.txt. And (3), we represent each tweet as a list of (ID, word count) tuples. E.g., assuming the following mapping (word:ID): {ill:0, explain:1, everything:2, in:3, a:4, …}
User ID,[(0, 2), (1, 2), (2, 2), (3, 1), (4, 1), … ]
word ID count
- tfidf Same as “count” except that instead of word counts, we provide the tfidf value as a measure of feature importance. E.g.,
User ID,[(0, 0.465), (1, 0.874), (2, 0.002), (3, 0.336), (4, 0.998), … ]
word ID tfidf
You can learn more about tfidf in Schutze et al.¨ (2008).
- glove300 We mapped each word to a 300-dimensional Glove “embedding vector”. These vectors were trained to capture the meaning of each word. We then average the vectors of each word in a tweet to obtain a single 300-dimensional representation of the tweet. E.g.,
User ID, 2.05549970e-02 8.67250003e-02 8.83460036e-02 -1.26217505e-01 1.31394998e-02 […]
a 300-dimensional list of numbers
You can consult Pennington et al. (2014) for more information about GloVe.
Tasks
Stage I
- Feature Engineering (optional)
The process of engineering, or selecting, features that are useful for discriminating among your target class set is inherently poorly-defined. Most machine learning assumes that the attributes are simply given, with no indication from where they came. The question as to which features are the best ones to use is ultimately an empirical one: just use the set that allows you to correctly classify the data.
In practice, the researcher uses their knowledge about the problem to select and construct “good” features. What aspects of a tweet itself might indicate a user’s location? You can find ideas in published papers, e.g., Cheng et al. (2010).
We have discussed three types of attributes in this subject: categorical, ordinal, and numerical. All three types can be constructed for the given data. Some machine learning architectures prefer numerical attributes (e.g. kNN); some work better with categorical attributes (e.g. multivariate Naive Bayes) – you will probably observe this through your experiments.
It is optional for you to engineer some attributes based on the full Tweets dataset (and possibly use them instead of – or along with – the feature representations provided by us). Or, you may simply select features from the ones we generated for you (count, tfidf, and glove300).
- Machine Learning
Various machine learning techniques have been (or will be) discussed in this subject (Naive Bayes, Decision Trees, 0-R, etc.); many more exist. You are strongly encouraged to make use of machine learning software and/or existing libraries in your attempts at this project (such as sklearn or scipy).
The objective of your learners will be to predict the classes of unseen data. We will use a holdout strategy: the data collection has been split into three parts: a training set, a development set, and a test set. This data will be available on the LMS.
- The training phase will involve training your classifier and parameter tuning where required.
- The development phase is where you observe the performance of the classifier. The development data is labelled: you should run the classifier that you built in the training phase on this data to calculate one or more evaluation metrics to discuss and compare in your report, using tables/diagrams.
- The testing phase: The test data is unlabeled; you should use your preferred model to produce a prediction for each test instance, and submit your predictions to Kaggle website; we will use this output to confirm the observations of your approach.
To give you the possibility of evaluating your models on the test set, we will be setting up a Kaggle InClass competition. You can submit results on the test set there, and get immediate feedback on your system’s performance. There is a Leaderboard, that will allow you to see how well you are doing as compared to other classmates participating on-line. assumptions to
You should minimally implement and analyse in your report one baseline, and at least two different machine learning models. N.B. We are more interested in your critical analysis of methods and results, than the raw performance of your models.
